Is sandboxing sufficient to contain rogue agents?
Matthew Green argues that recent lab incidents do not yet prove that sandboxes are inherently ineffective: they do show serious containment and organizational failures. But useful training and evaluation agents need tools, data, and often network access, so stronger isolation alone cannot solve the problem. He adds a third risk: prompt injection can make otherwise compliant agents carry unauthorized instructions between shared systems.
The piece weighs the security view (labs have not implemented or governed containment well) against the alignment view (access needs and capable agents may defeat containment), and concludes that sandboxing helps but leaves a hard monitoring and authorization problem. In the Hacker News thread, some favor independent “warden” agents while others question whether that just moves the trust problem; another points out that generated code may still execute outside the sandbox. The blog’s three comments include both a claim that lax security is strategic and a rebuttal that breaches bring legal and regulatory costs; a separate commenter suggests a faster, specialized warden model.